The big AI companies are going to see their margins disappear

OPINION The future of AI is unwritten, but the writing is on the wall - your margin is my opportunity. Amazon founder Jeff Bezos said as much more than a decade ago in support of the e-souk's low-price, low-margin sales strategy. That opportunity exists in the AI training and inference business. But perhaps not for long. Two leading American AI companies, Anthropic and OpenAI, are not actually profitable at this point, but their pitch to investors is something along the lines of "just hang in there a few more years and keep sending cash." Given reports that Claude Code subscribers paying $200 a month can potentially consume $5,000 worth of tokens and that OpenAI is also losing money on subscriptions, it starts to become a bit clear why Anthropic, OpenAI, Google, and Microsoft have already started pushing customers toward metered usage pricing. AI revenue needs to go up for frontier model makers to survive. And then AI adoption needs to grow. Government agencies and large corporations that don't keep a close eye on fees may be terrified enough of AI-enabled exploitation to pay a premium for models like Anthropic's Mythos and OpenAI's GPT-5.5. But more price-sensitive folk may shop for cheaper tokens. And they're likely to find them. Benedict Evans, among the more astute industry observers, expects AI models will be commoditized. In his recently updated presentation, "AI eats the world," he suggests that the AI supply/demand imbalance will ease and the pricing power of leading AI labs will dissipate. He argues that models will become commodity infrastructure and that innovation and pricing power will have to move up the stack. That's already evident in Anthropic's efforts to keep developers interacting through its own tools like the Claude Code CLI and desktop app, and through services that sit atop its models like Claude Cowork, Claude Design, and Claude for Creative Work. But it's more apparent in US companies lobbying for regulatory intervention as a defense against competition from China, some of which has taken the form of copying AI models via a process called distillation. Zilan Qian, a research associate at the Oxford China Policy Lab, recently explored how software developers in China are acquiring AI tokens for pennies on the dollar. She writes that despite the fact that leading US model makers try to prevent people in China from using US models, everyone who wants access can get it through API proxies. "The logs they generate may have become a commodity, traded for purposes ranging from model training to targeted fraud," Qian wrote. "Meanwhile, every layer of control frontier US AI companies have added (geoblocking, phone verification, credit card requirements, and now live biometric KYC checks) has produced a corresponding layer of evasion infrastructure." This process may not be savory or sustainable - Qian posits these token sellers are just trying to acquire customers and obtain data - but it points to the difficulty US firms will have maintaining their margins and their exclusivity. Open weight models like GLM-5.1, Kimi K2.6, DeepSeek V4-Pro, and Qwen3-Coder-Next are already adequate for less demanding software development work and some, like Qwen3.6-27B, run quite well on suitably provisioned local hardware. US companies are estimated to have a lead of about seven months on Chinese AI companies. But that race will not go on forever. Even if US AI models continue to improve at their current pace, open weight models from China and elsewhere should match current leaders Claude Opus 4.7 and OpenAI GPT-5.5 by the end of 2026. At that point, better benchmarks will no doubt be welcomed, but they won't be necessary. Commodity AI will be good enough for enterprise and entrepreneurial software development. And maybe other uses will emerge, but coding right now is what people are paying for. As noted by Andreessen Horowitz, annualized AI spending by enterprises reached $3 billion annually for coding. In other categories (legal $500 million, support $400M, and medical/health $300M), adoption is significantly less. Looking at Evans's "AI eats the world" figures, promoting AI adoption will be a challenge. The tech industry is the only US workplace sector where more than 25 percent use AI on a daily basis. In finance, professional services, healthcare, retail, manufacturing, and government, there's less daily usage. And in the consumer space, only five percent of ChatGPT's 900 million-plus weekly users pay for the privilege. Among software developers, most of those using AI are not trying to apply it to cutting-edge research or to develop complex attack chains. They're using it for fairly well understood software applications and workflows, or they're experimenting with AI agents. And increasingly, it looks like they can buy tokens at a discount if that matters. Anthropic and OpenAI need pricing and adoption to go up in order to thrive. Their margin is their vulnerability. They're going to strike deals with incumbents to make their models available on desktop and mobile hardware, particularly given the space and power constraints of phones. That will come at a cost. The likely winners will be the companies that control software distribution and delivery - operating system vendors like Apple, Google, and Microsoft, and cloud service providers like Amazon, Google, and Microsoft. Absent regulatory or legal barriers, supply constraints, or practical obstacles, prices face downward pressure where margins are high. And when you're many billions in the hole like Anthropic and OpenAI, that makes escape more difficult. In his presentation, Evans observes, "Sometimes software eats the world, and sometimes it only nibbles." (R)